The judged dimensions are readings of the item, and on a borderline one the verdict can differ between requests. AuditorConfig(votes=N) audits each sample N times and returns the majority verdict per dimension. It configures every audit stage: the Auditor, and the audit and reaudit stages of a Refinery. The fix stage has no votes; its own control is FixerConfig(max_attempts=N).
votes=1, the default, is a single audit. Use an odd number: a tie has no majority, so an even count is refused.
On the HTTP API the same setting is config: {"votes": 3} on /audit and config: {"audit": {"votes": 3}, "reaudit": {"votes": 1}} on /refine.
Cost
Each vote is an audit call, billed per token on the work actually done. Three votes on the audit before the fix and one on the reaudit is a reasonable split: the first decides what the fix changes, the second confirms the result.
Why it matters for fixes
Inside Refinery, the audit before the fix decides what the fix changes and the reaudit says whether it helped. When the audit is noisy, a fix can be routed by a flip on the original item rather than a real defect. Voting the audit stage takes that noise out of the routing; voting the reaudit makes the proof stable too. Last modified on September 18, 2026
